In view of the scattered straw mulching in the field, the various straw shapes, the difficulty in identifying the fine straw, and the traditional image recognition methods are disturbed by factors such as light and shadow easily. Taking Longjiang County, Qiqihar City, Heilongjiang Province as the research site, a field straw image dataset was constructed. After cropping and labeling the image, a straw detection model based on U-Net network was constructed. Changing the network structure of the coding stage to the first four layers of ResNet34 as the feature extractor, the complexity of the model was increased and the extraction of straw features was enhanced. In order to enhance the detailed identification of straw edges, the multibranch asymmetric dilated convolutional block (MADC Block) was used to extract multi-scale image features on the deep feature map at the highest semantic information layer. In order to increase the detection ability of fine straws, dense feature pyramid networks (DFPN) were used in the skip connection stage to perform information fusion of low-level feature maps and high-level feature maps. Using the feature map to correspond to the difference of the receptive fields in the straw image, the problem of variety of straw shapes was solved. In order to avoid the invalid calculation of straw feature map during upsampling, the decoding stage used fast up-convolution block (FUC Block) was used for upsampling. Experiments result showed that the average intersection ratio of the algorithm on the straw image dataset collected by the vehicle camera was 84.78%, which was 2.59 percentage points higher than that of U-Net. The average processing time of the network for images with a size of 640 pixels×480 pixels was less than 3ms. Compared with manual measurement, the error was less than 5%, which met the time complexity requirements of operation detection. The algorithm can improve the identification of straw in the shadow area to a certain extent, and improve the identification ability of fine straw.